Deep learning models now can achieve expert-level performance for chest radiographic (CXR) diagnosis, but they require pretraining with large-scale image datasets and categorical labels. Such labels are, however, frequently incomplete and commonly fail to capture the full spectrum of clinical knowledge embedded in radiology reports issued by expert radiologists. Large language models (LLMs) can encode text into machine-understandable semantic representations, providing an opportunity for cross-modal knowledge distillation, thereby enhancing vision-based CXR diagnosis models. We propose CREED—Classifying and Reconstructing Expanded Embeddings for cross-modal knowledge Distillation to address this shortcoming. CREED has three learning objectives: (1) embedding reconstruction to preserve fine-grained language information, (2) diagnostic classification to bridge the modality gap, and (3) KL-divergence minimization to enforce alignment between vision and language embeddings. Extensive experiments across diverse diagnostic tasks and on report generation demonstrate that CREED consistently outperforms baseline and state-of-the-art (SOTA) fully-/self-supervised models. The results highlight the effectiveness of cross-modal knowledge distillation from LLMs for enhancing CXR diagnosis by infusing clinically rich semantics into vision models, ultimately improving diagnostic accuracy, interpretability, and generalization across diverse medical imaging tasks. The codes and pretrained models are available at GitHub.com/JLiangLab/CREED .